Digital online AI application processing platform and optimization method

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a digital online AI application processing platform and optimization method, which comprises the following steps of: comprehensively analyzing an AI model architecture of edge equipment, constructing and optimizing a low-dimensional manifold, and improving the operation efficiency of the edge equipment; by collecting AI model information of multiple edge devices, a high-dimensional manifold of a distributed cloud platform is constructed, and centralized management and dynamic updating of the information are achieved; the model performance is greatly optimized by calculating the tangent space of the low-dimensional manifold of the AI model of the edge device; the consistency of the model is improved by extracting a model difference vector and identifying a generality adjustment demand; by introducing Riemannian measurement, the difference between models is measured, and the cooperative work capability is enhanced; and by identifying similar model groups and integrating tasks, reasonable allocation of the tasks and optimal configuration of resources are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and particularly relates to a digital online AI application processing platform and an optimization method. Background Art

[0002] Currently, there are still some challenges in AI application processing, including the following aspects: The AI models lack flexibility and adaptability. Most existing AI platforms are optimized based on predefined model structures and parameters, lacking the ability to automatically adjust the AI model structure according to the characteristics of input data and runtime performance feedback, making the AI models appear less flexible and adaptable when dealing with complex and changing application scenarios; The combination problem of edge computing and cloud computing. Currently, AI platforms focus on centralized cloud computing or separately explore edge computing, but how to effectively combine the powerful computing power of cloud computing and the low latency and high reliability advantages of edge computing to achieve more efficient AI application processing is still a problem; The dynamic update and maintenance of AI models are insufficient. Existing AI platforms have deficiencies in the dynamic update and maintenance of models and are difficult to quickly respond to changing requirements. For this reason, the present invention proposes a digital online AI application processing platform and an optimization method. Summary of the Invention

[0003] The purpose of the present invention is to solve the problems in the background art and propose a digital online AI application processing platform and an optimization method.

[0004] To achieve the above purpose, the present invention adopts the following technical solutions: A digital online AI application processing platform and an optimization method, including: S1. Comprehensively analyze the AI model architecture carried by edge devices; comprehensively evaluate the edge device resource capabilities and AI task requirements, and construct and optimize the low-dimensional manifold based on the edge device AI model; S2. Collect AI model information of multiple edge devices through establishing a communication mechanism, compare and analyze the commonalities and differences, and use the manifold learning algorithm to construct a high-dimensional manifold based on the distributed cloud platform. At the same time, perform dynamic update and maintenance; S3. Use differential geometry methods to calculate the tangent space of the low-dimensional manifold of the edge device AI model, and perform tangent space analysis and vector relationship analysis. Based on the analysis results, formulate and implement an adjustment plan for the edge device AI model; S4. Perform tangent space calculation on the high-dimensional manifold of the distributed cloud platform, extract the model difference vector to represent the local model change direction on the low-dimensional manifold; by comparing the model difference vectors of multiple edge devices, mine similar optimization directions, identify common adjustment requirements; formulate corresponding adjustment strategies according to the common adjustment requirements; S5. Define the Riemannian metric on the low-dimensional manifold of the edge device AI model to measure the model differences, evaluate the distance between the adjustment schemes of multiple edge device AI models and the target model structure, and determine the optimal adjustment scheme for the edge device AI model; S6. Use the Riemannian metric to evaluate the structural differences of edge device AI models in the distributed cloud platform, identify similar model groups through clustering, integrate tasks based on the similar groups and dynamically allocate them to achieve collaborative work of edge devices.

[0005] Furthermore, conduct a comprehensive analysis of the AI model architecture carried by the edge device; comprehensively evaluate the resource capabilities of the edge device and the requirements of AI tasks. The process of constructing and optimizing the low-dimensional manifold based on the edge device AI model includes: Record the types of each layer of the AI model carried by the edge device, as well as the connection methods between layers; count the number of neurons and parameter dimension information of each layer; extract the structural features and parameter distribution characteristics of each layer of the edge device AI model; By evaluating the computing resources of the edge device, analyze the matching degree between the requirements of the AI application task and the computing resources to obtain the resource limit analysis result; According to the resource limit analysis result, combined with the structural features and parameter distribution characteristics of the edge device AI model, determine the dimension parameters of the low-dimensional manifold; use the manifold embedding technology to map each layer of the edge device AI model into the low-dimensional manifold subspace to form a low-dimensional representation; during the mapping process, accurately map the key information of the weights of neurons and activation function parameters to the corresponding coordinates on the low-dimensional manifold; According to the initial structure of the low-dimensional manifold, adjust its curvature and topological structure to be more approximate to the local model structure: the initial structure of the low-dimensional manifold is obtained through the manifold embedding technology based on the structural features and parameter distribution characteristics of the edge device AI model; for curvature adjustment, the curvature is used to reflect the degree of bending of the low-dimensional manifold of the edge device AI model at any manifold feature point, that is, in the low-dimensional manifold of the edge device AI model, the size and distribution of the curvature match the actual structure of the model; by analyzing the connection methods and parameter distribution between each layer of the edge device AI model, identify the regions for adjusting the curvature; for topological structure adjustment, the topological structure describes the connection relationship between each manifold node in the manifold, that is, in the low-dimensional manifold of the edge device AI model, the topological structure is used to reflect the hierarchical relationship and interaction method between each layer of the edge device AI model; if the topological structure of the initial manifold does not match the actual structure of the edge device AI model, adjust it by adding or reducing nodes in the manifold, changing the connection method between nodes, or introducing new topological features.

[0006] Furthermore, the process of collecting AI model information of multiple edge devices through establishing a communication mechanism, comparing and analyzing the commonalities and differences, and constructing a high-dimensional manifold based on a distributed cloud platform using manifold learning algorithms, and simultaneously performing dynamic update and maintenance includes: Establish stable communication connections with multiple edge devices for real-time collection of AI model structure information of each edge device; obtain the AI model structure information of each edge device, including the parameter values, connection status, and activation function types of each layer; integrate the AI model structure information from different edge devices to generate a comprehensive information set of the edge device AI model structure; Compare and analyze the structures of AI models of different edge devices, analyze their common architectural features and device-specific differences, and identify the general structural features with universality and the customized structural differences reflecting individual differences; According to the results of the commonality and difference analysis, determine the information range to be covered by the high-dimensional manifold, that is, the common information and difference information of AI models of different edge devices; use manifold learning algorithms to integrate all the AI model structure information of edge devices; based on all the AI model structure information of edge devices on the distributed cloud platform, construct a high-dimensional and complex manifold representation; reflect the actual distribution of the edge device AI model structure by adjusting the parameters and structure of the high-dimensional manifold; Establish a dynamic update mechanism for the high-dimensional manifold to monitor the changes in the edge device AI model structure and the requirements of new tasks in real time; according to the real-time monitoring results, timely adjust the parameters and structure of the high-dimensional manifold to adapt to the changing edge device AI model structure and new task requirements; formulate a maintenance strategy for the high-dimensional manifold.

[0007] Furthermore, the process of calculating the tangent space of the low-dimensional manifold of the edge device AI model using differential geometry methods, performing tangent space analysis and vector relationship analysis, and formulating and implementing an adjustment plan for the edge device AI model based on the analysis results includes: On the constructed low-dimensional manifold of the edge device AI model, select a point and mark this point as the current parameter configuration point; approximate the local change characteristics of the edge device AI model near this point by calculating the tangent space at the current parameter configuration point; Make a small perturbation to the parameters of the edge device AI model to obtain several tangent vectors; calculate the partial derivatives of the edge device AI model parameters with respect to the current parameter configuration point through automatic differentiation technology; based on the calculated partial derivatives, form a set of basis vectors of the tangent space, and this set of basis vectors is used to represent all possible small adjustment directions of the edge device AI model at the current parameter configuration point; Based on the tangent space of the low-dimensional manifold of the edge device AI model, the performance indicators of the edge device AI model are monitored in real time, and a performance threshold is set. When the model performance indicator is lower than the performance threshold, it is determined that the model performance has degraded, and the tangent space analysis is started: the vector directly related to the performance change of the edge device AI model is extracted from the tangent space, that is, the performance association vector, where the performance association vector represents the local performance change direction of the edge device AI model on the low-dimensional manifold; Use vector operations and statistical analysis methods to obtain the relationship between vectors and performance changes of edge device AI models: For vector operations, use the basis vectors in the tangent space to construct a set of vectors indirectly related to performance changes through linear combination; calculate the inner product or angle between these vectors and the performance-related vectors to quantify the correlation between them; for statistical analysis, collect multiple sets of edge device AI model performance change data and corresponding tangent vector data to form a data set associated with performance changes and tangent vectors; based on the data set associated with performance changes and tangent vectors, use the statistical method of regression analysis to find the quantitative relationship between the tangent vector and the performance change of the edge device AI model, that is, establish a regression model with the tangent vector as the independent variable and the performance change of the edge device AI model as the dependent variable, and analyze the linear or nonlinear relationship between them; based on the results of vector operations and statistical analysis, identify that the tangent vector is highly correlated with the performance change of the edge device AI model; Based on the results of tangent space and vector analysis, determine the direction and magnitude of the model structure adjustment, that is, whether to increase or decrease the number of neurons corresponding to the direction or adjust the type of activation function; and determine the magnitude of the structure adjustment based on the degree of change in edge device AI model performance and the size of the vector; at the same time, formulate an edge device AI model adjustment plan, including the type of adjusted parameters, adjustment magnitude, and adjustment sequence, and implement the adjustment on the edge device, while monitoring the performance changes of the adjusted model.

[0008] Furthermore, the tangent space calculation is performed on the high-dimensional manifold of the distributed cloud platform, and the model difference vector is extracted to represent the change direction of the local model on the low-dimensional manifold; by comparing the model difference vectors of multiple edge devices, similar optimization directions are mined and common adjustment requirements are identified; the process of formulating corresponding adjustment strategies according to common adjustment requirements includes: Perform tangent space calculations on the high-dimensional manifold of the distributed cloud platform and the low-dimensional manifold of each edge device AI model to obtain their respective tangent space representations: In the high-dimensional manifold of the distributed cloud platform, the high-dimensional manifold is composed of all edge device AI model structures, and each model structure state point corresponds to a structural state of an edge device AI model; for each model structure state point on the high-dimensional manifold, calculate the tangent space of the high-dimensional manifold by solving the Jacobian matrix of the manifold at that point, and represent the calculated tangent space in the form of a matrix or vector set; at the same time, obtain the tangent space of the low-dimensional manifold of the edge device AI model to obtain their respective tangent space representations; Extracting a vector from the tangent space of the low-dimensional manifold of each edge device AI model, namely, a model difference vector, wherein the model difference vector represents the direction of local model change of each edge device AI model on the low-dimensional manifold; Use the vector similarity measurement method to compare the vector directions of the low-dimensional manifold tangent space of the AI ​​models of multiple edge devices and explore similar optimization directions: extract the model difference vector from the tangent space of the low-dimensional manifold of each edge device AI model; use the cosine similarity measurement method to calculate the similarity between the model difference vectors of different edge devices, that is, measure the similarity of the directions of the difference vectors of the two models, with a value range of [-1,1]; based on the similarity calculation results, explore similar optimization directions and identify common adjustment needs; When it is found that multiple edge devices have common adjustment needs, the distributed cloud platform formulates corresponding adjustment strategies and pushes the common adjustment strategies to relevant edge devices through communication channels; based on the results of high-dimensional manifold tangent space analysis, the execution efficiency of different tasks on edge devices and distributed cloud platforms is evaluated.

[0009] Furthermore, the Riemann metric on the low-dimensional manifold of the edge device AI model is defined to measure the model difference, and the distance between the adjustment schemes of multiple edge device AI models and the target model structure is evaluated. The process of determining the optimal adjustment scheme of the edge device AI model includes: On the low-dimensional manifold of the edge device AI model, a Riemannian metric is defined to measure the distance between different model structures and parameters. Let the low-dimensional manifold of the edge device AI model be \(R\), and any two model structures be \(p, q\in R\). Then the Riemannian metric \(d(p, q)\) represents the distance between \(p\) and \(q\). When the edge device adjusts the model, multiple possible adjustment schemes for the edge device AI model are generated, and each scheme corresponds to different model structure and parameter changes. The Riemannian metric is used to calculate the distance between each adjustment scheme and the target model structure. Let the target model structure be \(t\in R\), and the adjustment scheme be \(F_i\in R\), where \(i\) is the adjustment scheme index and \(i = 1, 2,\cdots\). Then the distance \(d(F_i, t)\) represents the difference between the scheme \(F_i\) and the target \(t\). According to the calculated distances between each adjustment scheme and the target model structure, the superiority and inferiority order of each adjustment scheme are evaluated. Considering the resource limitations of the edge device, the adjustment schemes are screened, and the schemes whose resource requirements exceed the tolerance of the edge device are eliminated. Among the remaining adjustment schemes that meet the resource requirements after screening, the scheme with the closest distance to the target model structure under the Riemannian metric and meeting the resource requirements is further selected as the optimal adjustment scheme. The optimal adjustment scheme of the edge device AI model is implemented on the edge device, and a new model structure is deployed. After implementing the adjustment scheme of the edge device AI model, the performance indicators of the edge device AI model are continuously monitored. According to the monitoring results, a decrease or abnormality in the model performance is promptly detected and fed back to the adjustment scheme selection process.

[0010] Furthermore, the process of using the Riemannian metric to evaluate the differences in the edge device AI model structures in the distributed cloud platform, identifying similar model groups through clustering, integrating tasks based on the similar groups, and dynamically allocating them to achieve the collaborative work of edge devices includes: On the high-dimensional manifold of the distributed cloud platform, a Riemannian metric is defined to measure the differences between different edge device AI model structures. Let the high-dimensional manifold of the distributed cloud platform be \(G\), and any two edge device AI model structures be \(P_j, Q_k\in G\), where \(j, k\) are the edge device indices. Then the Riemannian metric \(d(P_j, Q_k)\) represents the distance between \(P_j\) and \(Q_k\). The Riemannian metric is used to calculate the distances between each edge device AI model structure, and a distance matrix \(D\) is constructed. Where \(D[j, k]\) represents the distance \(d(P_j, Q_k)\) between the AI model structures of edge device \(j\) and edge device \(k\). Based on the distance matrix D, the spectral clustering combined with the local consistency method is used to analyze the similarity between the AI model structures of each edge device: by taking the reciprocal of the distance S[j,k] = 1 / (1+D[j,k]), the distance matrix D is converted into a similarity matrix S; the element S[j,k] in the similarity matrix S represents the similarity between the AI model structures of edge device j and edge device k; construct a local similarity matrix L, for each edge device j, find the edge devices in its neighborhood or near neighbors, and calculate the similarity between them and edge device j; the element L[j,k] in the local similarity matrix L represents the local similarity between edge device j and edge device k in its neighborhood, if edge device k is not in the neighborhood of edge device j, then L[j,k] = 0; the global similarity matrix S and the local similarity matrix L are weighted and fused to obtain the final similarity matrix M; Use the similarity matrix M for spectral clustering, and calculate the Laplacian matrix L(M) of M: L(M) = D(M) - M, where D(M) is the diagonal matrix of M; solve the eigenvalues and eigenvectors of the Laplacian matrix L(M); select the first N eigenvectors to form a new feature space; apply the K-means clustering algorithm on the new feature space for clustering, and assign a cluster label to the AI model structure of the edge device; According to the clustering results, the AI model structures of edge devices with the same cluster label are divided into the same similarity group, that is, the edge devices with similar structures in the edge device AI model are divided into the same similarity group; the edge devices within the similarity group are assigned to the same task group for task integration, so that the edge devices within the task group can cooperate to complete the same task or a part of the task; according to the task requirements and the resource status of the edge devices, the tasks are assigned to each task group; during the task execution process, continuously monitor the change of the model structure using the Riemannian metric to timely detect the abnormality of the edge device AI model structure; at the same time, monitor the task execution deviation; according to the monitoring results, timely adjust the task assignment and model integration strategies, including re-dividing the similarity group, adjusting the task assignment plan, and optimizing the edge device AI model structure.

[0011] Compared with the prior art, the beneficial effects of the present invention are: by comprehensively analyzing the edge device AI model architecture and comprehensively evaluating the resource capabilities and task requirements, effectively improving the running adaptability of the AI model; by establishing a communication mechanism to collect the AI model information of multiple edge devices, realizing the centralized management of the AI model and improving the management efficiency; by using the differential geometry method to calculate the tangent space and implementing the targeted adjustment plan, significantly optimizing the model performance; by constructing a high-dimensional manifold based on the distributed cloud platform, realizing the dynamic update and maintenance of the model and effectively enhancing the timeliness of the model; by using the Riemannian metric to evaluate the model structure differences and identifying the similar model groups, promoting the collaborative work of edge devices and improving the overall efficiency. Description of the Drawings

[0012] Figure 1 This is a flowchart of a digital online AI application processing platform and optimization method proposed by the present invention. Specific implementation manners

[0013] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0014] Refer to Figure 1 , a digital online AI application processing platform and optimization method, including: S1. Comprehensively analyze the AI model architecture carried by the edge device; comprehensively evaluate the resource capabilities of the edge device and the AI task requirements, and construct and optimize the low-dimensional manifold based on the edge device AI model; wherein, the edge device AI model refers to the AI model running in the edge computing environment; S2. Collect AI model information of multiple edge devices through establishing a communication mechanism, compare and analyze the commonalities and differences, and use the manifold learning algorithm to construct a high-dimensional manifold based on the distributed cloud platform. At the same time, perform dynamic update and maintenance; S3. Use differential geometry methods to calculate the tangent space of the low-dimensional manifold of the edge device AI model, and perform tangent space analysis and vector relationship analysis. Based on the analysis results, formulate and implement an adjustment plan for the edge device AI model; S4. Perform tangent space calculation on the high-dimensional manifold of the distributed cloud platform, extract the model difference vector to represent the local model change direction on the low-dimensional manifold; by comparing the model difference vectors of multiple edge devices, mine similar optimization directions, identify common adjustment requirements; formulate corresponding adjustment strategies according to the common adjustment requirements; S5. Define the Riemannian metric on the low-dimensional manifold of the edge device AI model to measure the model difference, evaluate the distance between the adjustment plans of multiple edge device AI models and the target model structure, and determine the optimal adjustment plan for the edge device AI model; wherein, the target model structure refers to an ideal or optimized model structure and parameter configuration that the edge device AI model is expected to achieve; S6. Use the Riemannian metric to evaluate the structural differences of the edge device AI models in the distributed cloud platform, identify similar model groups through clustering, integrate tasks based on the similar groups and dynamically allocate them to achieve collaborative work of the edge devices.

[0015] It should be further noted that in the specific implementation process, a comprehensive analysis of the AI model architecture carried by the edge device is conducted; the process of comprehensively evaluating the resource capabilities of the edge device and the requirements of AI tasks and constructing and optimizing the low-dimensional manifold based on the edge device AI model is as follows: Record the types of each layer of the AI model carried by the edge device, including fully connected layers, convolutional layers, pooling layers, and the connection methods between layers, namely sequential connection and skip connection; count the number of neurons and parameter dimension information of each layer, including the size of the weight matrix and the length of the bias vector; extract the structural features and parameter distribution characteristics of each layer of the edge device AI model to provide a basis for subsequent low-dimensional manifold construction: traverse each layer of the edge device AI model, identify and record the type of each layer, including fully connected layers, convolutional layers, pooling layers, and activation layers; analyze and record the connection methods between layers, namely sequential connection (the output of the previous layer is used as the input of the next layer) and skip connection (cross-layer connection, such as the residual connection in the residual network); obtain the number of neurons in each layer, obtain the size of the weight matrix of the fully connected layer (the number of input neurons multiplied by the number of output neurons) and the length of its bias vector (the number of output neurons), obtain the number of convolutional kernels, size, stride, and the length of its bias vector (the number of convolutional kernels) in the convolutional layer, obtain the pooling window size and stride of the pooling layer; record the type of activation function used in each layer; by calculating the basic statistics (mean, median, mode, variance, standard deviation, etc.) of the parameters of each layer of the edge device AI model, obtain the parameter distribution characteristics, that is, the central tendency and dispersion degree of the parameters; By evaluating the computing resources of edge devices (CPU / GPU performance, storage capacity), analyzing the matching degree between AI application task requirements and computing resources, the resource limit analysis results are obtained: for the CPU / GPU performance and storage capacity of computing resources, they are respectively defined as fuzzy variables, including three performance fuzzy sets of high performance, medium performance, and low performance, and three capacity fuzzy sets of large capacity, medium capacity, and small capacity; membership functions are defined for each fuzzy variable of computing resources to describe the degree to which each resource level belongs to different fuzzy sets; the membership functions are used to convert the actual resource values (such as CPU speed, memory size) into fuzzy quantities; the real-time performance and accuracy of AI application task requirements are defined as fuzzy variables, obtaining three real-time performance fuzzy sets of high real-time performance, medium real-time performance, and low real-time performance; the accuracy is divided into three accuracy fuzzy sets of high accuracy, medium accuracy, and low accuracy; membership functions corresponding to each fuzzy variable of AI application task requirements are defined to describe the degree to which the task requirements belong to different fuzzy sets; fuzzy rules are established to define the matching relationship between resource levels and task requirements, and the matching degree is defined as multiple fuzzy sets, including high matching degree, medium matching degree, and low matching degree. For example, if the CPU performance is high performance and the storage capacity is large capacity, then the matching degree between the resource level and the task requirements is high; if the real-time performance requirement is high real-time performance and the accuracy requirement is medium accuracy, then the matching degree between the resource level and the task requirements depends on the size of the storage capacity; the fuzzy matching degree is converted into a clear value through the maximum membership degree method for subsequent analysis; during the operation of edge devices, continuously monitor the changes in computing resources and AI application task requirements, and dynamically adjust the fuzzy rules and membership functions to adapt to new situations; According to the resource limit analysis results, combined with the structural characteristics and parameter distribution characteristics of the edge device AI model, determine the dimensional parameters of the low-dimensional manifold to meet the resource requirements of the edge device while ensuring the approximate accuracy of the model; use manifold embedding techniques, such as Locally Linear Embedding (LLE), Isometric Mapping (ISOMAP), to map each layer of the edge device AI model (including neuron weights, activation function parameters) into the low-dimensional manifold subspace to form a low-dimensional representation; during the mapping process, accurately map the key information of the neuron weights and activation function parameters to the corresponding coordinates on the low-dimensional manifold; Adjust the curvature and topological structure of the low-dimensional manifold according to its initial structure to approximate the local model structure more closely: The initial structure of the low-dimensional manifold is obtained through manifold embedding technology based on the structural characteristics and parameter distribution characteristics of the edge device AI model; for curvature adjustment, curvature is used to reflect the degree of bending of the low-dimensional manifold of the edge device AI model at any manifold feature point, that is, in the low-dimensional manifold of the edge device AI model, the magnitude and distribution of curvature match the actual structure of the model; by analyzing the connection method and parameter distribution between the layers of the edge device AI model, the area for adjusting curvature is identified. For example, if there are complex non-linear relationships in the edge device AI model, the corresponding manifold area is represented with a larger curvature; for topological structure adjustment, the topological structure describes the connection relationship between the manifold nodes in the manifold, that is, in the low-dimensional manifold of the edge device AI model, the topological structure is used to reflect the hierarchical relationship and interaction method between the layers of the edge device AI model; if the topological structure of the initial manifold does not match the actual structure of the edge device AI model, it is adjusted by adding or reducing nodes in the manifold, changing the connection method between nodes, or introducing new topological features. For example, if there are skip connections or residual connections in the edge device AI model, the corresponding topological structure is introduced into the initial manifold to represent these connections; during the optimization process, ensure that the low-dimensional manifold meets the model approximation accuracy without exceeding the resource limitations of the edge device.

[0016] It should be further noted that in the specific implementation process, the process of collecting the information of multiple edge device AI models through establishing a communication mechanism, comparing and analyzing the commonalities and differences, and constructing a high-dimensional manifold based on a distributed cloud platform using manifold learning algorithms, and at the same time, performing dynamic update and maintenance is as follows: Establish stable communication connections with multiple edge devices for real-time collection of the structure information of each edge device AI model; obtain the structure information of each edge device AI model, including the parameter values, connection status, and activation function types of each layer; integrate the AI model structure information from different edge devices to generate a comprehensive information set of the edge device AI model structure for subsequent high-dimensional manifold construction; Conduct a comparative analysis of the structures of different edge device AI models, analyze their common architectural features and device-specific differences, and identify universal common structural features and customized structural differences that reflect individual differences. Among them, common architectural features refer to the structural characteristics that commonly exist in different edge device AI models. For example, most models use a combination of convolutional layers and pooling layers to extract features and use ReLU as the common activation function. Device-specific differences refer to the model structure differences caused by different hardware conditions or application scenarios of different edge devices. Universal common structural features are the concretization of common architectural features, specifically referring to the structural characteristics that widely exist and have a certain universality in different models, such as the standard convolutional-pooling stacking mode. Customized structural differences are the specific manifestations of device-specific differences, specifically the structural characteristics that reflect the individual differences of different devices, such as a unique network module of a certain device or optimization adjustments for specific hardware. According to the results of the commonality and difference analysis, determine the information range that the high-dimensional manifold needs to cover, that is, the common information and difference information of different edge device AI models, to reflect the diversity of the structures of edge device AI models. Use manifold learning algorithms, such as principal component analysis (PCA), t-SNE, etc., to integrate all the structural information of edge device AI models, including the number of layers, the type of each layer (such as fully connected layer, convolutional layer, etc.), the number of neurons, and the parameter dimensions. Based on all the structural information of edge device AI models on the distributed cloud platform, construct a high-dimensional and complex manifold representation. By adjusting the parameters and structure of the high-dimensional manifold, reflect the actual distribution of the structures of edge device AI models. Establish a dynamic update mechanism for the high-dimensional manifold to monitor the changes in the structures of edge device AI models and the requirements of new tasks in real time. According to the real-time monitoring results, timely adjust the parameters and structure of the high-dimensional manifold to adapt to the continuously changing structures of edge device AI models and new task requirements. Develop a maintenance strategy for the high-dimensional manifold, including regular updates, exception handling, etc., to ensure the stability and reliability of the high-dimensional manifold.

[0017] It should be further noted that in the specific implementation process, the process of calculating the tangent space of the low-dimensional manifold of the edge device AI model using differential geometry methods, conducting tangent space analysis and vector relationship analysis, and formulating and implementing an adjustment plan for the edge device AI model based on the analysis results is as follows: On the constructed low-dimensional manifold of the edge device AI model, select a point and mark this point as the current parameter configuration point. Among them, the current parameter configuration point represents the specific parameter state or configuration of the edge device AI model at a certain moment. By calculating the tangent space at the current parameter configuration point, approximately represent the local change characteristics of the edge device AI model near this point. Perform a small perturbation on the AI model parameters of the edge device to obtain a number of tangent vectors; through the automatic differentiation technique, calculate the partial derivatives of the AI model parameters of the edge device with respect to the current parameter configuration point; based on the calculated partial derivatives, form a set of basis vectors for the tangent space, and this set of basis vectors is used to represent all possible small adjustment directions of the edge device AI model at the current parameter configuration point; According to the tangent space of the low-dimensional manifold of the edge device AI model, monitor the performance metrics of the edge device AI model in real time, such as accuracy, recall rate, etc., and set a performance threshold. When the model performance metric is lower than the performance threshold, it is determined that the model performance has declined, and the tangent space analysis is started: extract the vectors directly related to the performance change of the edge device AI model from the tangent space, that is, the performance correlation vectors, where the performance correlation vectors represent the local performance change direction of the edge device AI model on the low-dimensional manifold; Use vector operations and statistical analysis methods to obtain the relationship between the vectors and the performance change of the edge device AI model: for vector operations, use the basis vectors in the tangent space to construct a set of vectors indirectly related to the performance change through linear combination; calculate the inner product or angle between these vectors and the performance correlation vectors (that is, the vectors directly related to the model performance change extracted from the tangent space) to quantify their correlation; for statistical analysis, collect multiple sets of edge device AI model performance change data and corresponding tangent vector data to form a performance change and tangent vector correlation data set; based on the performance change and tangent vector correlation data set, use the statistical method of regression analysis to find the quantitative relationship between the tangent vectors and the performance change of the edge device AI model, that is, establish a regression model, with the tangent vectors as independent variables and the performance change of the edge device AI model as the dependent variable, and analyze their linear or non-linear relationship; comprehensively consider the results of vector operations and statistical analysis to identify that the tangent vectors are highly correlated with the performance change of the edge device AI model; According to the tangent space and vector analysis results, judge the direction and magnitude of the model structure adjustment, that is, whether to increase or decrease the number of neurons corresponding to this direction or adjust the type of activation function; and combine the degree of performance change of the edge device AI model and the vector magnitude to determine the structure adjustment magnitude, such as adjusting the number of neurons, the parameters of the activation function, etc.; at the same time, formulate an adjustment plan for the edge device AI model, including the type of parameters to be adjusted, the adjustment magnitude, and the adjustment order, and implement the adjustment on the edge device, and monitor the performance change of the adjusted model to evaluate the adjustment effect.

[0018] It should be further noted that in the specific implementation process, the process of calculating the tangent space of the high-dimensional manifold of the distributed cloud platform, extracting the model difference vectors to represent the local model change direction on the low-dimensional manifold, mining similar optimization directions by comparing the model difference vectors of multiple edge devices, identifying common adjustment requirements, and formulating corresponding adjustment strategies is as follows: Perform tangent space calculations on the high-dimensional manifold of the distributed cloud platform and the low-dimensional manifold of each edge device AI model to obtain their respective tangent space representations: In the high-dimensional manifold of the distributed cloud platform, the high-dimensional manifold is composed of all edge device AI model structures, and each model structure state point corresponds to the structural state of an edge device AI model; wherein the model structure state point represents the complete structural information of the edge device AI model at a certain moment or in a certain configuration, including the number of layers of the model, the type of each layer (such as fully connected layer, convolutional layer, etc.), the number of neurons, and the parameter dimension; for each model structure state point on the high-dimensional manifold (i.e., each model structure), calculate the tangent space of the high-dimensional manifold by solving the Jacobian matrix of the manifold at that point, and represent the calculated tangent space in the form of a matrix or vector set; at the same time, obtain the tangent space of the low-dimensional manifold of the edge device AI model to obtain their respective tangent space representations; Extracting a vector from the tangent space of the low-dimensional manifold of each edge device AI model, namely, a model difference vector, wherein the model difference vector represents the direction of local model change of each edge device AI model on the low-dimensional manifold; Use the vector similarity measurement method to compare the vector directions of the tangent space of the low-dimensional manifold of the AI ​​model of multiple edge devices, and mine similar optimization directions: extract the model difference vector from the tangent space of the low-dimensional manifold of each edge device AI model; use the cosine similarity measurement method to calculate the similarity between the model difference vectors of different edge devices, that is, measure the similarity of the directions of the two model difference vectors, with a value range of [-1,1]. The closer to 1, the more similar the directions are; based on the similarity calculation results, mine similar optimization directions and identify common adjustment needs. For example, if the model difference vectors of multiple edge devices are close to 1 in cosine similarity, it indicates that their local change directions on the low-dimensional manifold are similar, and there may be common optimization needs; When it is found that multiple edge devices have common adjustment needs, the distributed cloud platform will formulate corresponding adjustment strategies (such as uniformly adjusting the number of a certain type of neurons, the type of activation function, etc.), and push the common adjustment strategies to related edge devices through communication channels to achieve collaborative optimization; among them, the common adjustment needs indicate that some edge devices have commonalities in the direction of model optimization, and the tasks on the devices can be uniformly allocated or adjusted; according to the results of high-dimensional manifold tangent space analysis, the execution efficiency of different tasks on edge devices and distributed cloud platforms is evaluated (including computing resource utilization, communication overhead, task completion time, etc.). If it is found that there is a task that is more conducive to overall performance improvement when executed on an edge device, the task allocation strategy is adjusted to transfer the task from the distributed cloud platform to the edge device for execution. At the same time, the execution effect of the task after transfer is monitored to ensure that the overall performance is improved.

[0019] It should be further noted that in the specific implementation process, the process of defining the Riemannian metric on the low-dimensional manifold of the edge device AI model to measure the model difference, evaluating the distance between the adjustment schemes of multiple edge device AI models and the target model structure, and determining the optimal adjustment scheme of the edge device AI model is as follows: On the low-dimensional manifold of the edge device AI model, a Riemannian metric is defined to measure the distance between different model structures and parameters. The Riemannian metric can be constructed based on the spatial distribution of model parameters, gradient information, or other geometric characteristics to ensure that it can accurately reflect the differences between model structures. Let the low-dimensional manifold of the edge device AI model be \(R\), and any two model structures be \(p,q\in R\). Then the Riemannian metric \(d(p, q)\) represents the distance between \(p\) and \(q\). When the edge device performs model adjustment, multiple possible adjustment schemes are generated, and each scheme corresponds to different model structure and parameter changes. The Riemannian metric is used to calculate the distance between each adjustment scheme and the target model structure. Let the target model structure be \(t\in R\), and the adjustment scheme be \(F_i\in R\) (\(i\) is the adjustment scheme index, and \(i = 1, 2,\cdots\)). Then the distance \(d(F_i,t)\) represents the difference between the scheme \(F_i\) and the target \(t\). According to the calculated distances between each adjustment scheme and the target model structure, the pros and cons of each adjustment scheme are evaluated: the smaller the distance, the better the adjustment scheme and the closer it is to the ideal state. Considering the resource limitations of the edge device, the adjustment schemes are screened to eliminate the schemes whose resource requirements exceed the capacity of the edge device. Among the remaining adjustment schemes that meet the resource requirements after screening, the scheme with the closest distance to the target model structure under the Riemannian metric and meeting the resource requirements is further selected as the optimal adjustment scheme. The optimal adjustment scheme of the edge device AI model is implemented on the edge device, and the new model structure is deployed. After implementing the edge device AI model adjustment scheme, the performance indicators of the edge device AI model are continuously monitored. According to the monitoring results, a decrease or abnormality in the model performance is promptly detected and fed back to the adjustment scheme selection process for subsequent optimization and adjustment.

[0020] It should be further noted that in the specific implementation process, the process of using the Riemannian metric to evaluate the differences in the edge device AI model structures in the distributed cloud platform, identifying similar model groups through clustering, integrating tasks based on the similar groups, and dynamically allocating them to achieve the collaborative work of edge devices is as follows: On the high-dimensional manifold of the distributed cloud platform, a Riemannian metric is defined to measure the differences between different edge device AI model structures. Let the high-dimensional manifold of the distributed cloud platform be \(G\), and any two edge device AI model structures be \(P_j,Q_k\in G\), where \(j,k\) are the edge device indices. Then the Riemannian metric \(d(P_j,Q_k)\) represents the distance between \(P_j\) and \(Q_k\). The Riemannian metric is used to calculate the distances between the edge device AI model structures, and a distance matrix \(D\) is constructed. Among them, \(D[j,k]\) represents the distance \(d(P_j,Q_k)\) between the AI model structures of edge device \(j\) and edge device \(k\). Based on the distance matrix D, the spectral clustering combined with the local consistency method is used to analyze the similarity between the AI model structures of each edge device: by taking the reciprocal of the distance S[j,k] = 1 / (1+D[j,k]), the distance matrix D is converted into a similarity matrix S; the element S[j,k] in the similarity matrix S represents the similarity between the AI model structures of edge device j and edge device k; construct the local similarity matrix L, for each edge device j, find the edge devices in its neighborhood or vicinity, and calculate the similarity between them and edge device j; the element L[j,k] in the local similarity matrix L represents the local similarity between edge device j and edge device k in its neighborhood, if edge device k is not in the neighborhood of edge device j, then L[j,k] = 0; the global similarity matrix S and the local similarity matrix L are weighted and fused (M =αS +βL, where α and β are weight factors) to obtain the final similarity matrix M; Use the similarity matrix M for spectral clustering, calculate the Laplacian matrix L(M) of M: L(M) = D(M)-M, where D(M) is the diagonal matrix of M; solve the eigenvalues and eigenvectors of the Laplacian matrix L(M); select the first N eigenvectors (N is the desired number of clusters) to form a new feature space; apply the K-means clustering algorithm on the new feature space for clustering, and assign a cluster label to the AI model structure of the edge device; According to the clustering results, the AI model structures of edge devices with the same cluster label are divided into the same similarity group, that is, the edge devices with similar structures in the edge device AI model are divided into the same similarity group; the edge devices within the similarity group are assigned to the same task group for task integration, so that the edge devices within the task group cooperate to complete the same task or a part of the task; according to the task requirements and the resource status of the edge devices, the tasks are assigned to each task group; during the task execution process, continuously monitor the changes in the model structure using the Riemannian metric, and timely detect abnormalities in the edge device AI model structure, such as model drift, excessive parameter changes, etc.; at the same time, monitor the task execution deviation, such as extended task completion time, decreased resource utilization rate, etc.; according to the monitoring results, timely adjust the task assignment and model integration strategies, including re-dividing the similarity group, adjusting the task assignment plan, and optimizing the edge device AI model structure to improve the overall task execution efficiency.

[0021] It should be understood that in various embodiments of the present application, the magnitude of the serial numbers of the above processes does not mean the order of execution, and the order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0022] It should be understood that determining B based on A does not mean determining B only based on A, and B can also be determined based on A and / or other information.

[0023] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the said claims.

[0024] Finally: The above are only the preferred embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A digital online AI application processing platform and optimization method, characterized by: S1. Comprehensively analyze the AI ​​model architecture of edge devices; comprehensively evaluate the resource capabilities of edge devices and AI task requirements, and build and optimize low-dimensional manifolds based on edge device AI models; S2. Collect AI model information of multiple edge devices by establishing a communication mechanism, compare and analyze commonalities and differences, and use manifold learning algorithms to build high-dimensional manifolds based on distributed cloud platforms. At the same time, perform dynamic updates and maintenance. S3. Use differential geometry methods to calculate the tangent space of the low-dimensional manifold of the edge device AI model, and perform tangent space analysis and vector relationship analysis. Based on the analysis results, formulate and implement an edge device AI model adjustment plan; S4. Perform tangent space calculation on the high-dimensional manifold of the distributed cloud platform, extract the model difference vector to represent the local model change direction on the low-dimensional manifold; by comparing the model difference vectors of multiple edge devices, explore similar optimization directions and identify common adjustment requirements; formulate corresponding adjustment strategies based on the common adjustment requirements; S5. Define the Riemann metric on the low-dimensional manifold of the edge device AI model to measure the model difference, evaluate the distance between the adjustment schemes of multiple edge device AI models and the target model structure, and determine the optimal adjustment scheme of the edge device AI model; S6. Use Riemann metric to evaluate the differences in the AI ​​model structures of edge devices in the distributed cloud platform, identify similar model groups through clustering, integrate tasks based on similar groups and dynamically allocate them to achieve collaborative work of edge devices.

2. According to claim 1, a digital online AI application processing platform and optimization method, characterized in that: Comprehensively analyze the AI ​​model architecture of edge devices; comprehensively evaluate the resource capabilities of edge devices and AI task requirements, and build and optimize the low-dimensional manifold based on the edge device AI model. The process includes: Record the types of each layer of the AI ​​model carried by the edge device, as well as the connection methods between layers; count the number of neurons and parameter dimension information of each layer; and extract the structural characteristics and parameter distribution characteristics of each layer of the edge device AI model; By evaluating the computing resources of edge devices, analyzing the matching degree between AI application task requirements and computing resources, we can obtain the resource limitation analysis results. According to the results of resource limitation analysis, combined with the structural characteristics and parameter distribution characteristics of the edge device AI model, the dimension parameters of the low-dimensional manifold are determined; using manifold embedding technology, each layer of the edge device AI model is mapped to the low-dimensional manifold subspace to form a low-dimensional representation; in the mapping process, the key information of the neuron weights and activation function parameters are accurately mapped to the corresponding coordinates on the low-dimensional manifold; According to the initial structure of the low-dimensional manifold, its curvature and topological structure are adjusted to be more similar to the local model structure: the initial structure of the low-dimensional manifold is obtained through manifold embedding technology based on the structural characteristics and parameter distribution characteristics of the edge device AI model; for curvature adjustment, the curvature is used to reflect the degree of bending of the low-dimensional manifold of the edge device AI model at any manifold feature point, that is, in the low-dimensional manifold of the edge device AI model, the size and distribution of the curvature match the actual structure of the model; by analyzing the connection method and parameter distribution between the layers of the edge device AI model, the area for adjusting the curvature is identified; for topological structure adjustment, the topological structure describes the connection relationship between the manifold nodes in the manifold, that is, in the low-dimensional manifold of the edge device AI model, the topological structure is used to reflect the hierarchical relationship and interaction method between the layers of the edge device AI model; if the topological structure of the initial manifold does not match the actual structure of the edge device AI model, it is adjusted by adding or reducing the nodes in the manifold, changing the connection method between the nodes, or introducing new topological features.

3. According to claim 1, a digital online AI application processing platform and optimization method, characterized in that: By establishing a communication mechanism, we collect AI model information from multiple edge devices, compare and analyze commonalities and differences, and use manifold learning algorithms to build a high-dimensional manifold based on a distributed cloud platform. At the same time, the process of dynamic update and maintenance includes: Establish stable communication connections with multiple edge devices to collect AI model structure information of each edge device in real time; obtain AI model structure information of each edge device, including parameter values, connection status, and activation function type of each layer; integrate AI model structure information from different edge devices to generate a comprehensive information set of AI model structure of edge devices; Compare and analyze the structures of different edge device AI models, analyze their common architectural features and device-specific differences, and identify universal general structural features and customized structural differences that reflect individual differences; According to the results of commonality and difference analysis, determine the information scope that the high-dimensional manifold needs to cover, that is, the commonality information and difference information of different edge device AI models; use the manifold learning algorithm to integrate the structural information of all edge device AI models; build a high-dimensional and complex manifold representation based on the structural information of all edge device AI models on the distributed cloud platform; reflect the actual distribution of edge device AI model structures by adjusting the parameters and structure of the high-dimensional manifold; Establish a dynamic update mechanism for high-dimensional manifolds to monitor changes in the edge device AI model structure and new task requirements in real time; adjust the parameters and structure of the high-dimensional manifold in a timely manner based on real-time monitoring results to adapt to the ever-changing edge device AI model structure and new task requirements; and formulate a maintenance strategy for the high-dimensional manifold.

4. The digital online AI application processing platform and optimization method according to claim 1, characterized in that: The process of using differential geometry methods to calculate the tangent space of the low-dimensional manifold of the edge device AI model, and performing tangent space analysis and vector relationship analysis, and formulating and implementing the edge device AI model adjustment plan based on the analysis results includes: On the constructed low-dimensional manifold of the edge device AI model, select a point and mark it as the current parameter configuration point; by calculating the tangent space at the current parameter configuration point, approximately represent the local change characteristics of the edge device AI model near the point; Make slight perturbations to the edge device AI model parameters to obtain several tangent vectors; calculate the partial derivatives of the edge device AI model parameters with respect to the current parameter configuration point through automatic differentiation technology; construct a set of basis vectors in the tangent space based on the calculated partial derivatives, which are used to represent all possible slight adjustment directions of the edge device AI model at the current parameter configuration point; Based on the tangent space of the low-dimensional manifold of the edge device AI model, the performance indicators of the edge device AI model are monitored in real time, and a performance threshold is set. When the model performance indicator is lower than the performance threshold, it is determined that the model performance has degraded, and the tangent space analysis is started: the vector directly related to the performance change of the edge device AI model is extracted from the tangent space, that is, the performance association vector, where the performance association vector represents the local performance change direction of the edge device AI model on the low-dimensional manifold; Use vector operations and statistical analysis methods to obtain the relationship between vectors and performance changes of edge device AI models: For vector operations, use the basis vectors in the tangent space to construct a set of vectors indirectly related to performance changes through linear combination; calculate the inner product or angle between these vectors and the performance-related vectors to quantify the correlation between them; for statistical analysis, collect multiple sets of edge device AI model performance change data and corresponding tangent vector data to form a data set associated with performance changes and tangent vectors; based on the data set associated with performance changes and tangent vectors, use the statistical method of regression analysis to find the quantitative relationship between the tangent vector and the performance change of the edge device AI model, that is, establish a regression model with the tangent vector as the independent variable and the performance change of the edge device AI model as the dependent variable, and analyze the linear or nonlinear relationship between them; based on the results of vector operations and statistical analysis, identify that the tangent vector is highly correlated with the performance change of the edge device AI model; Based on the results of tangent space and vector analysis, determine the direction and magnitude of the model structure adjustment, that is, whether to increase or decrease the number of neurons corresponding to the direction or adjust the type of activation function; and determine the magnitude of the structure adjustment based on the degree of change in edge device AI model performance and the size of the vector; at the same time, formulate an edge device AI model adjustment plan, including the type of adjusted parameters, adjustment magnitude, and adjustment sequence, and implement the adjustment on the edge device, while monitoring the performance changes of the adjusted model.

5. The digital online AI application processing platform and optimization method according to claim 1, characterized in that: Perform tangent space calculations on the high-dimensional manifold of the distributed cloud platform and extract model difference vectors to represent the direction of local model changes on the low-dimensional manifold. By comparing the model difference vectors of multiple edge devices, similar optimization directions are discovered and common adjustment requirements are identified. The process of formulating corresponding adjustment strategies based on common adjustment needs includes: Perform tangent space calculations on the high-dimensional manifold of the distributed cloud platform and the low-dimensional manifold of each edge device AI model to obtain their respective tangent space representations: In the high-dimensional manifold of the distributed cloud platform, the high-dimensional manifold is composed of all edge device AI model structures, and each model structure state point corresponds to a structural state of an edge device AI model; for each model structure state point on the high-dimensional manifold, calculate the tangent space of the high-dimensional manifold by solving the Jacobian matrix of the manifold at that point, and represent the calculated tangent space in the form of a matrix or vector set; at the same time, obtain the tangent space of the low-dimensional manifold of the edge device AI model to obtain their respective tangent space representations; Extracting a vector from the tangent space of the low-dimensional manifold of each edge device AI model, namely, a model difference vector, wherein the model difference vector represents the direction of local model change of each edge device AI model on the low-dimensional manifold; Use the vector similarity measurement method to compare the vector directions of the low-dimensional manifold tangent space of the AI ​​models of multiple edge devices and explore similar optimization directions: extract the model difference vector from the tangent space of the low-dimensional manifold of each edge device AI model; use the cosine similarity measurement method to calculate the similarity between the model difference vectors of different edge devices, that is, measure the similarity of the directions of the difference vectors of the two models, with a value range of [-1,1]; based on the similarity calculation results, explore similar optimization directions and identify common adjustment needs; When it is found that multiple edge devices have common adjustment needs, the distributed cloud platform formulates corresponding adjustment strategies and pushes the common adjustment strategies to relevant edge devices through communication channels; based on the results of high-dimensional manifold tangent space analysis, the execution efficiency of different tasks on edge devices and distributed cloud platforms is evaluated.

6. The digital online AI application processing platform and optimization method according to claim 1, characterized in that: The process of defining the Riemann metric on the low-dimensional manifold of the edge device AI model to measure the model difference, evaluating the distance between the adjustment schemes of multiple edge device AI models and the target model structure, and determining the optimal adjustment scheme of the edge device AI model includes: On the low-dimensional manifold of the edge device AI model, the Riemann metric is defined to measure the distance between different model structures and parameters. Assuming the low-dimensional manifold of the edge device AI model is R, and any two model structures are p,q∈R, then the Riemann metric d(p, q) represents the distance between p and q; when the edge device adjusts the model, multiple possible edge device AI model adjustment schemes are generated, each corresponding to a different model structure and parameter changes; the Riemann metric is used to calculate the distance between each adjustment scheme and the target model structure; assuming that the target model structure is t∈R, the adjustment scheme is Fi∈R, where i is the adjustment scheme index, and i=1,2,..., then the distance d(Fi,t) represents the difference between the scheme Fi and the target t; according to the calculated distance between each adjustment scheme and the target model structure, the order of advantages and disadvantages of each adjustment scheme is evaluated; combined with the resource constraints of the edge device, the adjustment schemes are screened, and the schemes whose resource requirements exceed the bearing capacity of the edge device are eliminated; among the remaining adjustment schemes that meet the resource requirements after screening, the scheme that is closest to the target model structure and meets the resource requirements under the Riemann metric is further selected as the optimal adjustment scheme; the selected optimal adjustment scheme for the edge device AI model is implemented on the edge device, and the new model structure is deployed; after the edge device AI model adjustment scheme is implemented, the performance indicators of the edge device AI model are continuously monitored; according to the monitoring results, the model performance degradation or abnormality is discovered in time, and the feedback is fed back to the adjustment scheme selection process.

7. The digital online AI application processing platform and optimization method according to claim 1, characterized in that: The process of using Riemannian metric to evaluate the differences in the AI ​​model structures of edge devices in the distributed cloud platform, identifying similar model groups through clustering, integrating tasks based on similar groups and dynamically allocating them to achieve collaborative work of edge devices includes: On the high-dimensional manifold of the distributed cloud platform, the Riemann metric is defined to measure the differences between the AI ​​model structures of different edge devices. Let the high-dimensional manifold of the distributed cloud platform be G, and any two edge device AI model structures be Pj, Qk∈G, where j and k are edge device indexes, then the Riemann metric d(Pj,Qk) represents the distance between Pj and Qk. The Riemann metric is used to calculate the distance between the AI ​​model structures of each edge device and construct the distance matrix D. Where D[j,k] represents the distance d(Pj,Qk) between the AI ​​model structures of edge device j and edge device k. Based on the distance matrix D, spectral clustering combined with local consistency method is used to analyze the similarity between the AI ​​model structures of each edge device: by taking the inverse of the distance S[j,k] = 1 / (1+D[j,k]), the distance matrix D is converted into a similarity matrix S; the element S[j,k] in the similarity matrix S represents the similarity between the AI ​​model structures of edge device j and edge device k; the local similarity matrix L is constructed, and for each edge device j, its nearest neighbors or edge devices in the neighborhood are found, and the similarity between them and edge device j is calculated; the element L[j,k] in the local similarity matrix L represents the local similarity between edge device j and edge device k in its neighborhood. If edge device k is not in the neighborhood of edge device j, L[j,k] = 0; the global similarity matrix S and the local similarity matrix L are weightedly fused to obtain the final similarity matrix M; Use the similarity matrix M for spectral clustering and calculate the Laplacian matrix L(M) of M: L(M) = D(M)-M, where D(M) is the diagonal matrix of M; solve the eigenvalues ​​and eigenvectors of the Laplacian matrix L(M); select the first N eigenvectors to form a new feature space; apply the K-means clustering algorithm to the new feature space for clustering and assign a cluster label to the edge device AI model structure; According to the clustering results, the edge device AI model structures with the same cluster label are divided into the same similarity group, that is, the edge devices with similar structures in the edge device AI model are divided into the same similarity group; the edge devices in the similarity group are assigned to the same task group for task integration, so that the edge devices in the task group can collaboratively complete the same task or part of the task; tasks are assigned to each task group according to task requirements and the resource status of edge devices; during task execution, the Riemann metric is continuously used to monitor changes in the model structure to promptly discover abnormalities in the edge device AI model structure; at the same time, task execution deviations are monitored; according to the monitoring results, task allocation and model integration strategies are adjusted in a timely manner, including redividing similarity groups, adjusting task allocation plans, and optimizing the edge device AI model structure.

Images